Geração de imagens artificiais de vasos sanguíneos através de mapas de textura

• Ano de defesa: 2024
• Autor(a) principal: Carvalho, Adriano dos Reis
• Orientador(a): Comin, César Henrique
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Vasos sanguíneos; Modelo de vasos sanguíneos; Redes neurais; Geração de textura; Microscopia confocal; Estrutura tubular; Segmentação
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::MATEMATICA DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/20059

Images are present everywhere in modern daily life. Analyzing textures in images is highly relevant for identifying patterns and locating objects. When it comes to medical images of blood vessels, textural characteristics can aid in diagnoses and identifying the progression or regression of pathologies. Generally, literature works focus more on vessel geometry than on texture. It is common to consider vessels as primarily tubular structures with a Gaussian intensity profile. Our work aimed to fill this gap, showing the relevance of texture for the analysis of blood vessels. A methodology was developed for generating artificial images of blood vessels with textures extracted from real images. The method consists of generating vessel texture maps from manual annotations of real vessels. The maps are transformed to follow the geometry of randomly generated Bézier curves, enabling the generation of realistic vessel images. The potential of the methodology was demonstrated through pre-training neural networks on artificial images. It was verified that with only a few manually marked vessels, the networks achieve performance similar to the reference case where all vessels are marked. The developed method has the potential to reduce the manual annotation time needed to train neural networks.